An infrared insulator and fitting segmentation method and system based on improved DeepLabV3+

Abstract

The application discloses an infrared insulator and fitting segmentation method and system based on an improved DeepLabV3+. Firstly, an infrared image of a power transmission line is collected and a data set is constructed; secondly, a CBAM module is introduced in an encoding stage based on a DeepLabV3+ framework to enhance the expression of target salient features, Strip Pooling is integrated into an ASPP module to improve the context modeling capability of an elongated insulator structure, and an ECA module is embedded in a decoding stage to strengthen channel information interaction and boundary refinement; and finally, an insulator and fitting segmentation result is output. The method can effectively improve the accuracy, boundary continuity and robustness of infrared image segmentation under a complex background.

Description

Technical Field

[0001] This invention relates to the field of infrared image intelligent processing and transmission line fault detection technology, specifically to an infrared insulator and fitting segmentation method based on an improved DeepLabV3+. Background Technology

[0002] With the continuous development of intelligent inspection and condition monitoring technologies for power systems, infrared image-based transmission line equipment detection methods have become an important technical means for sensing the operational status of transmission lines due to their advantages such as non-contact operation, long-distance observation, and applicability to complex environments. Insulators and fittings, as key components in transmission lines, often have their surface thermal distribution, structural integrity, and abnormal conditions at connection points directly reflected in infrared images. Therefore, accurate segmentation of the insulator and fitting areas in infrared images is of great significance for subsequent fault identification, condition assessment, and intelligent operation and maintenance.

[0003] However, existing infrared insulator and fitting segmentation methods based on traditional image processing or conventional semantic segmentation networks still have significant limitations. On the one hand, infrared images of transmission lines have complex background components, often containing interference information such as the sky, towers, conductors, and environmental thermal radiation. Furthermore, insulators and fittings occupy a relatively small proportion of the overall image, easily leading to insufficient attention to the target area by the network. On the other hand, insulators in infrared images typically exhibit a long, thin, string-like distribution with blurred local boundaries and poor structural continuity. Fitting regions, on the other hand, are characterized by small size, irregular shape, and indistinct thermal features compared to adjacent regions. This results in existing methods being prone to problems such as target breakage, rough boundaries, loss of detail, and local missegmentation during the segmentation process. In addition, while the traditional DeepLabV3+ network has strong multi-scale feature extraction capabilities, it lacks sufficient feature enhancement capabilities for salient target areas in complex infrared scenes, does not adequately model long-range dependencies of long, thin structures, and its ability to recover local boundary details during the decoding stage needs improvement. Therefore, it is difficult to meet the practical application requirements for fine segmentation of infrared insulators and fittings in transmission lines. Summary of the Invention The purpose of this invention is to overcome the shortcomings of existing technologies and provide an improved infrared insulator and fitting segmentation method based on DeepLabV3+. By introducing the CBAM attention module, Strip Pooling module, and ECA module into the DeepLabV3+ network, the method enhances the network's segmentation capability for insulator and fitting regions from three aspects: target feature enhancement, elongated structure context modeling, and channel feature recalibration during the decoding stage. First, the infrared images of transmission lines are preprocessed to construct a segmentation dataset. Then, initial features of the input images are extracted using DeepLabV3+ as the basic framework. Next, the CBAM module is embedded in the encoding stage to enhance the network's attention to target regions. The Strip Pooling module is added to the ASPP module to improve the network's multi-scale context modeling capability for elongated insulator structures and fitting connection regions. The ECA module is introduced in the decoding stage to strengthen the channel expression capability after fusing high-level semantic features and low-level detail features. Finally, the segmentation results of the insulators and fittings are output, achieving automatic segmentation of target regions in complex infrared scenes.

[0004] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for segmenting infrared insulators and fittings based on an improved DeepLabV3+, the method comprising: S1: Collect infrared images of transmission lines, construct a segmentation dataset of insulators and fittings, and preprocess the input images to obtain the model input data; S2: Construct an infrared image segmentation network based on DeepLabV3+ framework, and extract initial features of the input image through the backbone feature extraction network; S3: In the encoding stage, the CBAM attention module is introduced to weight the initial features in the channel dimension and spatial dimension, enhance the feature expression of the target area of ​​insulators and fittings, and suppress complex background and irrelevant thermal radiation interference. S4: Introduce the Strip Pooling module into the ASPP module to perform strip pooling and multi-scale context fusion on the output features of the encoder, thereby enhancing the feature modeling capability for slender insulator structures and hardware connection areas. S5: In the decoding stage, an ECA module is introduced to enhance the channels of the fused high-level semantic features and low-level detail features, thereby improving the ability to express boundary regions and local detail features. S6: By upsampling and convolution classification, the segmentation results of insulators and fittings are output, realizing the automatic segmentation of the insulator region and fitting region in the infrared image of the transmission line.

[0005] Furthermore, in step S1, the input image is preprocessed, specifically including: performing size unification, normalization, and data enhancement on the acquired infrared image; the data enhancement includes one or more of random flipping, random rotation, brightness adjustment, and contrast adjustment; according to the target category in the image, the insulator region, hardware region, and background region are labeled at the pixel level to construct an infrared image segmentation dataset.

[0006] Furthermore, in S3, the execution steps of the CBAM attention module are as follows: global average pooling and global max pooling are performed on the feature map output by the backbone feature extraction network to generate channel attention weights; average pooling and max pooling are performed on the channel-enhanced feature map in the spatial dimension, and spatial attention weights are generated through convolution operation; the spatial attention weights are applied to the feature map to obtain the feature map after joint enhancement by channel attention and spatial attention.

[0007] Furthermore, in S4, the Strip Pooling module is set in the ASPP module, specifically: adding horizontal and vertical strip pooling branches on the basis of the original dilated convolution branches; extracting long-range dependency information in different directions through strip pooling operations; and fusing the output features of the strip pooling branches with the output features of convolution branches with different dilation rates to obtain enhanced multi-scale context features.

[0008] Furthermore, in S5, the execution steps of the ECA module are as follows: global average pooling is performed on the feature map after fusion in the decoding stage to obtain channel description information; one-dimensional convolution is used to model the local interaction relationship between each channel to generate channel weight coefficients; the channel weight coefficients are applied to the original feature map to obtain the enhanced feature map after channel recalibration.

[0009] Furthermore, in step S6, the specific steps for outputting the segmentation result images of insulators and fittings are as follows: the enhanced feature map is upsampled step by step to the original image resolution; the category of each pixel is predicted through a convolutional classification layer; and three segmentation result images of background, insulators, and fittings are generated based on the pixel classification results.

[0010] Furthermore, the method also includes a network training step, specifically: inputting the preprocessed infrared image and its corresponding labeled image into the improved DeepLabV3+ segmentation network; calculating the loss function based on the network output and the real labeled results; iteratively updating the network parameters using the backpropagation algorithm until the loss function converges, thereby obtaining the trained segmentation model.

[0011] Furthermore, the loss function includes the cross-entropy loss function and the Dice loss function, and the formula for calculating the total loss function is as follows:

[0012] Where Ltotal is the total loss function, Lce is the cross-entropy loss function, Ldice is the Dice loss function, and λ is the loss weight coefficient used to balance the proportion of the two types of losses during network training. Beneficial effects Compared with existing technologies, this infrared insulator and fitting segmentation method based on the improved DeepLabV3+ has the following advantages: This invention introduces a CBAM attention module into the DeepLabV3+ network, enabling the network to more effectively highlight the features of insulator and fitting target areas during the encoding stage, and suppressing the impact of complex backgrounds, irrelevant thermal radiation, and noise interference on the segmentation results. Simultaneously, a Strip Pooling module is added to the ASPP module, enhancing the network's ability to model long-range dependencies in slender insulator structures and fitting connection regions through directional strip pooling, effectively improving the problems of easy breakage of slender targets and unclear local connection relationships in traditional methods. Furthermore, an ECA module is embedded in the decoding stage to recalibrate the channels of the fused features, improving the ability to recover boundary details and discriminate local regions, thereby significantly improving the accuracy, boundary continuity, and model robustness of insulator and fitting segmentation in complex infrared scenes. Overall, this invention provides targeted improvements to address the problems of small targets, slender structures, complex backgrounds, and blurred boundaries in infrared transmission line images, and has strong engineering application value.

[0013] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0015] Figure 1 The flowchart shows the infrared insulator and fitting segmentation method based on the improved DeepLabV3+. Figure 2 This is a schematic diagram of the segmentation network structure of infrared insulators and fittings based on the improved DeepLabV3+. Detailed Implementation To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0016] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an improved infrared insulator and fitting segmentation method based on DeepLabV3+. By introducing the CBAM attention module, Strip Pooling module, and ECA module into the DeepLabV3+ network, the network's ability to extract features, perform multi-scale context modeling, and restore boundary details in infrared images of transmission lines is enhanced. This improves the problems of existing methods, such as insufficient focus on target areas in complex backgrounds, easy breakage of slender structures, easy omission of small local targets, blurred boundaries, and poor anti-interference ability. First, infrared images of transmission lines are acquired, filtered, preprocessed, and pixel-level labeled to construct a dataset suitable for infrared insulator and fitting segmentation. Then, a segmentation network based on DeepLabV3+ is constructed, and the initial features of the input image are extracted through the backbone feature extraction network. Next, the CBAM module is introduced in the encoding stage to enhance the response of the target salient region, and the Strip Pooling module is added to the ASPP module to enhance the long-range dependency information modeling capability of slender structures. In the decoding stage, the ECA module is introduced to improve the channel interaction and boundary detail recovery capabilities. Finally, the segmentation result images of insulators and fittings are output through upsampling and convolution classification, realizing automatic and fine segmentation of key components of transmission lines in complex infrared scenes.

[0017] The infrared insulator and fitting segmentation method based on the improved DeepLabV3+ described in this embodiment, such as Figure 1 As shown, the specific implementation steps are as follows: S1: Collect infrared images of transmission lines, construct a segmentation dataset of insulators and fittings, and preprocess the input images to obtain the model input data; Infrared images acquired during power transmission line inspections are collected. These images can originate from UAV inspection platforms, ground-based infrared thermal imaging equipment, online monitoring terminals, or manual inspection thermal imaging devices. Collection scenarios include inspection environments under clear, cloudy, nighttime, evening, and locally complex weather conditions. Due to the complex external operating environment of power transmission lines, the raw infrared images typically include not only insulators and hardware targets but also the sky background, towers, conductors, vegetation, and other thermal radiation areas. Furthermore, different images vary in terms of shooting distance, imaging angle, equipment resolution, and ambient temperature. Therefore, preprocessing of the raw data is necessary before model training.

[0018] First, the original infrared image samples are screened, removing invalid samples that are severely out of focus, have excessive target area occlusion, obvious noise stripes, severe thermal imaging distortion, or are difficult to identify insulator and hardware areas. Infrared images with complete target areas, clear thermal distribution characteristics, and representative scenes are retained. Pixel-level annotation is performed on the retained samples. According to the task requirements of this invention, image pixels are divided into three categories: background area, insulator area, and hardware area. The background area includes all areas except the target; the insulator area corresponds to the pixel area corresponding to the main body of the transmission line insulator string; and the hardware area corresponds to the pixel area corresponding to the connection between the insulator and the conductor and related hardware structures. A polygonal fine-grained annotation method is used during the annotation process to ensure that the boundary contour of the target area closely matches the actual structure, guaranteeing the accuracy of the training sample labels.

[0019] After annotation, all image samples are divided into training, validation, and test sets. To improve the model's generalization ability, the training set images undergo preprocessing and data augmentation. Specifically, the input images are uniformly scaled to a preset resolution and grayscale normalization is performed to reduce the impact of differences in grayscale distribution among different images. Simultaneously, the training images are augmented with random flipping, random rotation, brightness perturbation, contrast adjustment, random cropping, and the addition of slight noise to expand sample diversity and improve the model's adaptability to different backgrounds, targets of different scales, and different thermal distribution conditions. These processes yield the model input data used for subsequent network training and inference.

[0020] S2: Construct an infrared image segmentation network based on DeepLabV3+ framework, and extract initial features of the input image through the backbone feature extraction network; In this embodiment, the DeepLabV3+ network is used as the overall segmentation framework. The DeepLabV3+ network includes an encoder and a decoder. The encoder extracts high-level semantic features from the input image, while the decoder fuses these high-level semantic features with low-level detail features, restores the image's spatial resolution, and ultimately outputs pixel-level classification results. Because DeepLabV3+ features dilated convolutions and ASPP modules, it effectively balances receptive field expansion and multi-scale feature extraction, making it suitable as the foundational network structure for this invention.

[0021] After the preprocessed infrared image is input into the segmentation network, the initial features of the input image are first extracted by the backbone feature extraction network. The backbone feature extraction network can adopt a deep convolutional neural network structure, performing multi-layer convolution operations, normalization processing, and non-linear activation on the input image to gradually extract feature information at different levels. The feature map output by the front shallow network retains more edge, texture, and local hot spot information, which can better reflect the insulator outline and local details of the hardware; the feature map output by the rear deep network has stronger semantic expressive power and can characterize the semantic differences between the insulator and hardware regions and the complex background.

[0022] As the network depth increases, the spatial resolution of the feature maps gradually decreases, while the number of channels gradually increases. For insulator and fitting segmentation tasks, although deep features are beneficial for enhancing the semantic representation of the target, multiple downsampling can easily lead to the loss of boundary details of slender insulator structures and small-sized fitting regions, and irrelevant thermal radiation information in complex backgrounds may interfere with feature representation. Therefore, this invention further introduces CBAM, StripPooling, and ECA modules on the DeepLabV3+ framework to enhance key features at the encoding, context modeling, and decoding stages.

[0023] S3: In the encoding stage, the CBAM attention module is introduced to weight the initial features in the channel dimension and spatial dimension, enhance the feature expression of the target area of ​​insulators and fittings, and suppress complex background and irrelevant thermal radiation interference. During the encoding stage, the mid-to-high-level feature maps output by the backbone feature extraction network are input into the CBAM attention module. The CBAM attention module includes a channel attention submodule and a spatial attention submodule, which can adaptively weight the feature maps from the channel dimension and the spatial dimension, respectively, thereby improving the network's ability to focus on key target regions.

[0024] First, channel attention enhancement is performed on the input feature map. Specifically, global average pooling and global max pooling are performed on the input feature map to obtain two sets of feature vectors describing the channel response intensity. Global average pooling reflects the overall response level of each channel across the entire feature map, while global max pooling extracts the strongest activation response from each channel. These two sets of vectors are then fed into a shared multilayer perceptron for feature transformation, and an activation function is used to generate weight coefficients for each channel. Subsequently, these weight coefficients are multiplied channel-by-channel with the original feature map, giving higher weights to key channels related to insulators and fittings, while suppressing channel responses related to background, noise, and irrelevant heat sources, thus completing feature enhancement along the channel dimension.

[0025] After channel attention enhancement, spatial attention enhancement is performed. Specifically, average pooling and max pooling are performed along the channel dimension of the enhanced feature map to obtain two spatial response maps. These two maps are then concatenated along the channel dimension, and a spatial attention weight map is generated through convolution. This spatial attention weight map characterizes the importance of different spatial locations in the feature map. Multiplying this weight map pixel-by-pixel with the enhanced feature map further highlights the salient areas where insulators and fittings are located, while suppressing interference from background areas, sky thermal radiation, and irrelevant high-response areas.

[0026] By jointly enhancing channel attention and spatial attention through the CBAM module, the network can more accurately focus on insulator and fitting regions in infrared images during the encoding stage, improving the representation ability of target regions and providing more discriminative input features for subsequent multi-scale context modeling and boundary restoration. For infrared images with small target proportions and strong thermal radiation interference in complex backgrounds, this module can effectively reduce the adverse effects of background noise on feature extraction.

[0027] S4: Introduce the Strip Pooling module into the ASPP module to perform strip pooling and multi-scale context fusion on the output features of the encoder, thereby enhancing the feature modeling capability for slender insulator structures and hardware connection areas. After feature extraction and attention enhancement are completed in the encoding stage, the enhanced high-level feature map is input into the ASPP module. The ASPP module in traditional DeepLabV3+ uses multiple dilated convolution branches with different dilation rates, combined with a global pooling branch, to extract and fuse multi-scale contextual information. This structure performs well in general semantic segmentation tasks, but for the slender string structures of insulators and locally connected regions of fittings in infrared images of transmission lines, the conventional dilated convolution's ability to model long-range directional dependencies is still insufficient, easily leading to discontinuities in slender target structures or insufficient contextual association in local regions.

[0028] To address this, the present invention introduces a Strip Pooling module into the ASPP module. Based on the existing convolutional branches with different hole rates, horizontal and vertical strip pooling branches are added. The horizontal strip pooling branch extracts long-range dependencies of the target in the horizontal direction by aggregating feature map information along the horizontal direction; the vertical strip pooling branch extracts long-range dependencies of the target in the vertical direction by aggregating feature map information along the vertical direction. For infrared insulators, which typically exhibit a continuous, elongated string structure along a certain direction, strip pooling can better capture this continuous distribution feature, improving the network's ability to perceive the overall structural coherence. For fitting areas, although their area is small, they are often located at the connection point between the insulator and the conductor, and have a strong correlation with the surrounding structure; strip pooling can enhance the contextual relationship between this type of local region and its neighboring structure.

[0029] Specifically, the Strip Pooling module performs horizontal and vertical pooling operations on the input feature map to obtain directional awareness features. These features are then restored to their original size through convolutional mapping and upsampling, and finally concatenated and fused with the features output from each dilated convolutional branch in ASPP. In this way, the fused features not only contain the multi-scale semantic information extracted by traditional dilated convolution, but also additionally introduce long-range dependency features along the horizontal and vertical directions, thereby enhancing the network's ability to express the integrity of slender insulator structures, boundary continuity, and local relationships in the fitting connection region.

[0030] In this step, the ASPP output features enhanced by Strip Pooling have richer directional contextual information and stronger multi-scale expressive capabilities, providing more sufficient semantic support for detail recovery and target segmentation in the subsequent decoding stage.

[0031] S5: In the decoding stage, an ECA module is introduced to enhance the channels of the fused high-level semantic features and low-level detail features, thereby improving the ability to express boundary regions and local detail features. In the DeepLabV3+ decoding stage, high-level semantic features are upsampled and then fused with shallow low-level features to recover target boundaries and local details. For insulator and fitting segmentation tasks, although the target region features have been enhanced in the encoding and ASPP stages, issues such as blurred boundaries, weakened local details, and insufficient response to small targets may still occur during upsampling and feature fusion. To further improve the feature representation capability in the decoding stage, this invention introduces an ECA module to recalibrate the channels of the fused feature map.

[0032] Specifically, the feature map fused during the decoding stage is input into the ECA module. First, global average pooling is performed on the input feature map to obtain global statistical description information for each channel. This description information reflects the overall response level of each channel in the current feature map. Subsequently, one-dimensional convolution is used to model the local interaction relationships between channels, generating weight coefficients corresponding to each channel. Compared with traditional fully connected channel attention mechanisms, the ECA module does not require explicit dimensionality reduction, has a lighter structure, and can achieve effective cross-channel information interaction with lower computational cost. Finally, the channel weight coefficients are applied to the original fused feature map to enhance key channels and suppress redundant channels, thereby obtaining an enhanced feature map after channel recalibration.

[0033] After enhancement by the ECA module, the channel responses related to insulator edges, local hardware structures, and target details in the feature map during the decoding stage are further improved, enabling the network to better preserve target contours, boundary variations, and local feature differences during the restoration of spatial resolution. For regions with blurred boundaries and insignificant temperature differences between the target and background in infrared images, this module can effectively improve the feature separability of boundary and detail regions, mitigating issues such as rough contours, local missing features, and undersegmentation of small regions in the final segmentation results.

[0034] S6: By upsampling and convolution classification, the segmentation results of insulators and fittings are output, realizing the automatic segmentation of the insulator region and the fitting region in the infrared image of the transmission line; After channel enhancement in the decoding stage, the enhanced feature map is further upsampled step by step to restore a spatial resolution close to that of the original input image. Then, a convolutional classification layer is used to predict the class of each pixel, outputting the classification probability of each pixel belonging to the background, insulator, or hardware class. The class corresponding to the maximum classification probability at each pixel location is taken as the final predicted label, thus generating a complete pixel-level segmentation result map.

[0035] In the resulting segmentation image, different categories of regions can be visualized using different colors to facilitate subsequent inspection analysis and fault diagnosis. The background region corresponds to non-target parts, the insulator region corresponds to the main body of the insulator string, and the hardware region corresponds to the hardware connection parts connected to the insulator. Using this method, the automatic extraction of key component regions from infrared images of transmission lines can be achieved, providing fundamental data support for subsequent fault identification, temperature anomaly analysis, condition assessment, and intelligent inspection decision-making.

[0036] During network training, training set images and their corresponding labeled images are input into the improved DeepLabV3+ segmentation network. A loss function is calculated based on the difference between the network's predicted output and the true labels, and the network parameters are iteratively updated using a backpropagation algorithm. The loss function can be a weighted combination of cross-entropy loss and Dice loss, where cross-entropy loss constrains pixel classification accuracy, and Dice loss improves the overlap of class regions to enhance training performance under imbalanced sample conditions. Through multiple rounds of iterative training, the network gradually learns the discrimination features of insulator and fitting regions in different scenarios until the loss function converges or the validation set performance stabilizes, resulting in a successfully trained segmentation model.

[0037] This embodiment implements the infrared insulator and hardware segmentation method of the present invention for infrared inspection scenarios of transmission lines, as detailed below: First, infrared images acquired during transmission line inspections were collected, and insulator, hardware, and background regions were annotated at the pixel level to construct an infrared image segmentation dataset. Subsequently, the images underwent size unification, grayscale normalization, and data augmentation, and the processed data was divided into training, validation, and test sets. Next, the training images were input into an improved DeepLabV3+ segmentation network. Initial image features were extracted through a backbone feature extraction network. In the encoding stage, a CBAM module was introduced to enhance target feature response; in the ASPP module, a Strip Pooling module was introduced to enhance the contextual modeling capability of slender structures; and in the decoding stage, an ECA module was introduced to improve boundary recovery and local detail representation. After network training, a segmentation model suitable for transmission line infrared images was obtained.

[0038] During the testing phase, the infrared image to be segmented is input into the trained model. After feature extraction, attention enhancement, multi-scale context fusion, decoding recovery, and convolutional classification, the model outputs segmented images of insulators and fittings. For transmission line infrared image scenarios with complex backgrounds, low contrast, high requirements for the continuity of slender structures, and obvious local features of small targets, this embodiment can effectively improve the accuracy, boundary continuity, and overall robustness of insulator and fitting region segmentation.

[0039] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for segmenting infrared insulators and fittings based on an improved DeepLabV3+, characterized in that, The method includes: S1: Collect infrared images of transmission lines, construct a segmentation dataset of insulators and fittings, and preprocess the input images to obtain the model input data; S2: Construct an infrared image segmentation network based on DeepLabV3+ framework, and extract initial features of the input image through the backbone feature extraction network; S3: In the encoding stage, the CBAM attention module is introduced to weight the initial features in the channel dimension and spatial dimension, enhance the feature expression of the target area of ​​insulators and fittings, and suppress complex background and irrelevant thermal radiation interference. S4: Introduce the Strip Pooling module into the ASPP module to perform strip pooling and multi-scale context fusion on the output features of the encoder, thereby enhancing the feature modeling capability for slender insulator structures and hardware connection areas. S5: In the decoding stage, an ECA module is introduced to enhance the channels of the fused high-level semantic features and low-level detail features, thereby improving the ability to express boundary regions and local detail features. S6: By upsampling and convolution classification, the segmentation results of insulators and fittings are output, realizing the automatic segmentation of the insulator region and fitting region in the infrared image of the transmission line.

2. The infrared insulator and fitting segmentation method based on the improved DeepLabV3+ according to claim 1, characterized in that, In step S1, the input image is preprocessed, specifically including: The acquired infrared images are subjected to size unification, normalization, and data augmentation. The data enhancement includes one or more of the following: random flipping, random rotation, brightness adjustment, and contrast adjustment; Based on the target category in the image, pixel-level annotations are performed on the insulator region, hardware region, and background region to construct an infrared image segmentation dataset.

3. The infrared insulator and fitting segmentation method based on the improved DeepLabV3+ according to claim 1, characterized in that, In S3, the execution steps of the CBAM attention module are as follows: Global average pooling and global max pooling are performed on the feature maps output by the backbone feature extraction network to generate channel attention weights. The enhanced feature maps are subjected to average pooling and max pooling in the spatial dimension, and spatial attention weights are generated through convolution operations. Spatial attention weights are applied to the feature map to obtain a feature map that has been jointly enhanced by channel attention and spatial attention.

4. The infrared insulator and fitting segmentation method based on the improved DeepLabV3+ according to claim 1, characterized in that, In S4, the Strip Pooling module is set within the ASPP module, specifically as follows: Based on the existing dilated convolution branch, horizontal strip pooling branch and vertical strip pooling branch are added; Long-range dependency information in different directions is extracted through strip pooling operations; By fusing the output features of the strip pooling branch with the output features of the convolution branch with different dilation rates, we obtain the enhanced multi-scale context features.

5. The infrared insulator and fitting segmentation method based on the improved DeepLabV3+ according to claim 1, characterized in that, In step S5, the execution steps of the ECA module are as follows: Global average pooling is performed on the feature maps fused in the decoding stage to obtain channel description information; One-dimensional convolution is used to model the local interaction relationships between channels and generate channel weight coefficients; The channel weight coefficients are applied to the original feature map to obtain an enhanced feature map after channel recalibration.

6. The infrared insulator and fitting segmentation method based on the improved DeepLabV3+ according to claim 1, characterized in that, In step S6, the specific steps for creating the output insulator and fittings segmentation result diagram are as follows: The enhanced feature maps are upsampled step by step to the original image resolution; Each pixel is predicted using a convolutional classification layer; Based on the pixel classification results, three segmentation result images are generated: background, insulator, and hardware.

7. The infrared insulator and fitting segmentation method based on the improved DeepLabV3+ according to claim 1, characterized in that, The method also includes a network training step, specifically: The preprocessed infrared image and its corresponding labeled image are input into the improved DeepLabV3+ segmentation network; The loss function is calculated based on the network output and the actual labeled results. The backpropagation algorithm is used to iteratively update the network parameters until the loss function converges, thus obtaining the trained segmentation model.

8. The infrared insulator and fitting segmentation method based on the improved DeepLabV3+ according to claim 7, characterized in that, The loss function includes the cross-entropy loss function and the Dice loss function, and the formula for calculating the total loss function is as follows: Where Ltotal is the total loss function, Lce is the cross-entropy loss function, Ldice is the Dice loss function, and λ is the loss weight coefficient, used to balance the proportion of the two types of losses in network training.

9. An infrared insulator and fitting segmentation system based on an improved DeepLabV3+, characterized in that, include: The data acquisition and preprocessing module is used to acquire infrared images of transmission lines and construct a segmented dataset of insulators and fittings. The feature extraction module is used to extract features from the input image based on the DeepLabV3+ framework. The CBAM enhancement module is used to enhance the feature representation of the target region and suppress background interference; The Strip Pooling enhancement module is used to enhance the contextual modeling capabilities of slender structures and connected regions. The ECA enhancement module is used to improve the channel feature representation capability during the decoding stage. The segmentation output module is used to output the segmentation result diagram of insulators and fittings.

Images